ABEJA and Murata Manufacturing demonstrated a dual-arm robot running a vision-language-action model on real hardware, announced August 31, 2026. The model was built on NVIDIA’s Isaac GR00T N1.7 and trained by imitation learning from teleoperation data. In the demonstration the robot grips an object with the left arm, hands it to the right arm, reorients it and inserts it into a rack, in a test environment. The stated target is automating laboratory and experimental tasks.

Why the hand-off is the interesting part

Single-arm pick-and-place is close to solved in constrained settings. An arm-to-arm hand-off is not, because it requires the policy to hold a consistent estimate of the object through a moment when neither arm has full control of it, and then re-grasp in a pose chosen for the next step rather than the current one. Reorientation before insertion adds the same problem again: the grasp that gets the object out of the bin is rarely the grasp that gets it into the rack.

Doing this from imitation learning on teleoperation data, rather than scripted motion, is the claim being made. It is a company demonstration in a test environment, not a peer-reviewed result or a production deployment, and no success rate or cycle time was published.

Key Facts

  • Announced August 31, 2026 by ABEJA, jointly with Murata Manufacturing
  • Dual-arm robot running a VLA model built on NVIDIA Isaac GR00T N1.7
  • Trained by imitation learning from teleoperation data
  • Task sequence: left-arm grip → hand-off to right arm → reorientation → insertion into a rack, on real hardware in a test environment
  • Target application: automating laboratory and experimental work

Why it matters

Japanese industrial firms have generally approached physical AI through their own control stacks. Building on GR00T N1.7 is the other choice: take an open robot foundation model, supply your own hardware and your own teleoperation data, and compete on the task rather than the base model. It is the same division of labour J-HRTI’s data factory assumes — foreign or shared model weights, domestic data and domestic deployment.

Laboratory automation is also a shrewd target. Bench work is dexterous, repetitive, performed in a fixed and well-lit space, and valuable enough per hour to carry a robot’s cost — without the throughput demands that make sim-to-real gaps expensive on a production line.

Frequently Asked

What did ABEJA and Murata demonstrate?

A dual-arm robot running a vision-language-action model that grips an object with one arm, hands it to the other, reorients it and inserts it into a rack, on real hardware in a test environment. It was announced on August 31, 2026.

What model was it built on?

NVIDIA's Isaac GR00T N1.7, with the policy trained by imitation learning from teleoperation data.

What is it for?

The companies describe the target as automating laboratory and experimental tasks.

Was a success rate published?

No. The demonstration was reported without published success rates or cycle times, and took place in a test environment rather than a production deployment.

Sources & Further Reading